Fundamentals of Bioinformatics - Ogurtsov A.N. 2013

Information Principles in Biotechnology
Protein Analysis and Prediction
Protein Prediction Software

There are numerous prediction programs designed to identify unknown Proteins based on known chemical and Physical Properties of Amino Acids. Many of these programs are available online at the following links:

- http://expasy.org/Proteomics - via the ExPASy server, maintained by the Swiss Institute of Bioinformatics;

- http://www.embl.de/services/bioinformatics/index.php - via the EMBL servers at http://www.embl.fr/ or http://www.embl.de/, operated by the European Molecular Biology Laboratory;

- http://www.sanger.ac.uk/resources/software/ - via the Wellcome Trust Sanger Institute server;

- http://www.bioinf.manchester.ac.uk/dbbrowser/PRINTS/PRINTS.html - via the University of Manchester server;

- http://www.cmpharm.ucsf.edu/~nomi/nnpredict.html - a Protein Secondary Structure prediction program;

- http://pbil.ibcp.fr/htm/index.php?page=pbil_ibcp_Software.html - via the server of the Institute for the Biology and Chemistry of Proteins (IBCP);

- http://www.ch.embnet.org/software/COILS_form.html - a program for predicting coiled-coil regions within proteins;

- http://www.ebi.ac.uk/Tools/ - via the European Bioinformatics Institute (EBI) server;

- http://roselab.jhu.edu/dist/manual/index.html - the LINUS (Local Independently Nucleated Units of Structure) program.

The LINUS (Local Independently Nucleated Units of Structure) program is an ab initio Cell/13.html">Protein Structure Prediction tool based on the Amino Acid Sequence, developed by G.D. Rose and R. Srinivasan. This Procedure is entirely de novo, meaning it relies solely on the sequence itself without incorporating experimental data or known structural correlations. LINUS employs a "hierarchical" algorithm—folding begins with short peptide fragments and progressively combines them into longer ones.

The core concept behind LINUS is that the local structure of protein segments—short amino acid fragments—is determined by local interactions within those segments. During folding, each segment preferentially adopts the most energetically favorable Conformations. However, these preferred conformations, and even the most favorable ones that inevitably occur in the native protein structure, only slightly exceed the threshold of thermodynamic stability.

The local structure undergoes numerous transitions until a suitable stabilizing interaction is found. The computer favors local structures that appear more frequently during calculations and can thus pass their structural properties on to subsequent generations, thereby influencing the final outcome. The procedure uses a ratchet-like technical principle to guide computations along the most efficient path.

LINUS starts with an extended polypeptide chain of amino acids. During the calculation, randomly selected triplets of consecutive residues are perturbed, and the energy of the resulting conformation is evaluated. Sterically hindered structures are discarded, while remaining energy contributions are computed exclusively for local interactions.

To decide whether to accept a perturbed structure or revert to its predecessor, the Monte Carlo Method is used (see below). In LINUS, these steps are repeated many times, building up statistics on structural preferences for all amino acids.

Next, local fragments are assembled into larger units based on the calculated structural preference statistics. Throughout this process, the BOUNDARIES OF THE interaction-calculation zone continually expand—from local segments all the way to the entire globule.

The representation of the protein folding process in LINUS is fundamentally realistic, albeit approximate. The simulation includes all protein atoms except hydrogens, while the energy function is approximated and the dynamics are simplified. The energy functional accounts for:

1) steric repulsion between atoms;

2) the packing of buried hydrophobic residues;

3) Hydrogen Bonds;

4) salt bridges.

In most cases, LINUS successfully predicts the structures of small protein fragments (ranging from Supersecondary structures to domains) and can occasionally assemble them into the correct overall fold.

Monte Carlo Methods are widely used in protein structure calculations for efficient conformational sampling, minimizing complex Functions, and solving various other optimization tasks. Simple Minimization Procedures based on energy gradient descent are ineffective because calculations often become trapped in local minima far removed from the native structure.

The underlying idea of the method is to use random numbers to address problems that are difficult to compute analytically. The method's name was coined by one of its creators, Nicholas Constantine Metropolis, in honor of his uncle, who was an avid gambler.

To find the minimum of a multi-variable function using the Monte Carlo method—for instance, a protein energy function dependent on the variables that define its conformation—we assume that this function depends on a set of variables x, and for any given set of these variables, we can calculate the value of the function, which represents the energy of the protein conformation ε(x).

Here, x can be a set of protein atomic coordinates or the torsional angles of the main and side amino acid chains.

The Monte Carlo method is implemented According to the following procedure.

1. Initial conditions—a set of variables x—are randomly generated. The energy of the resulting conformation is calculated as ε = ε(х).

2. A perturbation of the variable x→x' is introduced to generate the next state.

3. The energy of the new state ε(х') is calculated.

4. A decision is made whether to accept the new perturbed conformation or revert to the previous initial one:

a) if the energy decreases, i.e., ε = ε(х) > ε(х'), the new state is accepted as the starting point for the next calculation step, with the new initial conditions set as x'→х and ε = ε(х').

b) if the energy increases or remains unchanged, i.e., ε(х) ≤ ε(х'), the new state can either be rejected or accepted with a probability of exp[-∆/(kT)], where ∆ = ε(х') - ε(х); k is the Boltzmann constant; and T is the effective Temperature.

5. Return to step 2.

The core feature of the method lies in step 4(b), as it allows the system to escape local minimum "traps".

The probability of accepting such an upward move in energy is determined by the effective temperature T. In this context, T is not a physical temperature, but rather a numerical parameter that controls the calculation.

Regardless of the temperature value, the probability of accepting a new structure with a higher energy decreases as this energy difference increases.

For any given value of ε, if the temperature is low, the value of ε(x)/(kT) will be high, whereas the value of exp[-ε(x)/(kT)] will be low.

Conversely, if the temperature is high, the value of ε(х)/(kT) will be low, and the exponential term exp[-ε(х)/(kT)] will approach unity.

As the temperature increases, so does the probability that a perturbed structure with a higher energy will be accepted.

This fairly simple idea has proven to be extremely effective for numerous Applications in protein structure modeling and other fields.

An extension of the Monte Carlo method exists—known as simulated annealing—in which the parameter T is varied: the temperature is initially set high to allow diverse conformational changes, and is subsequently lowered to drive the system toward an energy minimum.